Mot effektiv träning av kvantklassiska hybridmaskininlärningsmodeller : Profileringsstyrd optimering och praktisk parallellisering

Quantum computing provides potential for various scientific domains, especially through Quantum Machine Learning. Hybrid quantum-classical machine learning models are of great interests as they combine quantum capabilities with classical techniques to solving complex scientific and engineering problems. The Quantum Fourier Neural Operators, which is the quantum analog of classical Fourier Neural Operators, particularly highlight the potential of Quantum Machine Learning for efficiently solving Partial Differential Equations. Despite the theoretical advantages of these models, their practical application faces significant barriers due to the limitations of current quantum technologies, including the fact that quantum hardware development is still immature, poor scalability, and the high computational costs associated with quantum circuit execution. These limitations prevent their adoption and application in real-world scientific tasks. This thesis focus on the Partitioned Hybrid Quantum Fourier Neural Operator model training, addressing these challenges by introducing targeted strategies to enhance both the computational efficiency and scalability of hybrid quantum-classical models through two primary methodologies: profiling-guided optimization and parallelization techniques. Initially, a systematic profiling approach is applied to comprehensively identify performance bottlenecks across hybrid quantum-classical workflows. Based on insights derived from profiling, specific algorithmic optimizations are introduced. Supplementing the algorithmic optimizations, practical parallelization methods, including model parallelism and data parallelism, were also implemented to exploit computational independence and concurrency. Both parallelisms were explored using a Message Passing Interface based implementation. Model parallelism was also explored using the Dask-based approach. Experimental evaluations demonstrate the effectiveness of these combined approaches, showcasing significant improvements in overall training efficiency. Profiling-guided algorithmic optimizations alone resulted in approximately 20% reductions in execution time. Moreover, the Message Passing Interface based data parallelization along with optimizations yielded more than 70% reductions in training time, further demonstrating robust scalability with minor model performance tradeoff. The generality of the parallelization approach is further evaluated by applying it to a lighter-weight hybrid neural network model, confirming its broader applicability across hybrid Quantum Machine Learning architectures.